A prediction method for the refining properties of pure components based on hierarchical group construction

Abstract

The present invention discloses a prediction method for pure component refining properties based on hierarchical group construction. The method predicts the octane number and cetane number of pure component compounds in petroleum products. The hierarchical group construction is introduced through the hierarchical construction of component feature sets to avoid redundancy in feature sets. When the third-level component descriptor is added into the feature set, the Bayesian rule is introduced, so that the posterior probability distribution of the feature set can be estimated, and the feature set with higher posterior probability is selected, instead of focusing only on the prediction accuracy. On this basis, the Bayesian rule is introduced again, so that the posterior probability distribution of the final model can be estimated, avoiding the overfitting risk of the final prediction model. The invention can be applied to the crude oil and product blending unit in the petrochemical industry to effectively improve the accuracy of petroleum refining.

Description

Technical Field

[0001] The present invention relates to the field of data analysis of the refining industry, and in particular to a method for predicting the refining properties of pure components based on hierarchical group construction. Background Art

[0002] Traditional refining unit models, due to the limitations of analytical chemistry and computer hardware, mostly use lumped kinetic models, and raw materials and products are often divided into lumps based on macroscopic properties such as boiling point or solubility. For example, the ten-lumped and eleven-lumped models are widely used in catalytic cracking units. However, the lumped models based on macroscopic level division naturally have multi-component properties and cannot characterize component information in detail, making it difficult to extend such lumped models to new raw materials and catalyst systems. However, the lumped models at the molecular level can calculate the composition and properties of raw materials from the pure component level, establish reaction networks, and then accurately predict the properties of the products of the refining and processing units. Combined with the pure component property prediction model and the mixing rule model, the molecular dynamics model can not only predict the distribution of the refining unit products, that is, qualitative analysis, but also realize quantitative prediction of the corresponding refining properties of the products. This progress can enable decision makers to design the chemical structure of the pure components in the products in a targeted manner, optimize the unit operating conditions, and thus guide the direction of refining theory research and industrial production. The accuracy of the prediction of the refining properties of pure components is directly related to the accuracy of the product quality assessment, which in turn affects the optimization direction of each operating unit. It is the key point of the molecular dynamics model and is related to whether the molecular management technology can be successfully applied to refinery optimization. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a prediction method for the refining properties of pure components based on hierarchical group construction, and to predict the octane number of pure components in petroleum products, such as the octane number of each component in gasoline products and the cetane number of each component in diesel products. The component feature set is hierarchically constructed, and when the component descriptors are added to form a feature set, the Bayesian rule is introduced so that the posterior probability distribution can be estimated. On this basis, the hierarchical group construction is introduced, and the group fragments are hierarchically constructed to avoid the risk of overfitting in the final prediction.

[0004] The present invention specifically comprises the following steps:

[0005] Step 10, using the codable simplified component expression SMILES, the complex component structure is represented by two-dimensional coding, and a predefined group fragment component library is constructed, including primary groups, secondary groups and tertiary groups; the primary groups are basic groups containing component structures; the secondary groups are the link position combinations of the basic groups, which are used to distinguish aromatics from alkanes and the corresponding isomers; the tertiary groups are descriptors that describe the component topological structure;

[0006] Step 20, according to the molecular structure of the target component, a primary group and a secondary group are screened out from the predefined group fragment component library, and then according to the refining property to be predicted of the target component, a plurality of tertiary groups are screened out by using the principle of maintaining the minimum correlation with the primary group and the secondary group and maintaining the maximum amount of information with the property to be predicted, and any number of tertiary groups are randomly selected to form a plurality of component feature sets with the primary group and the secondary group, and then the component feature set with the largest posterior probability is screened out;

[0007] Step 30, using a linear cumulative function to combine groups at different levels for modeling and then solving the coefficients through the training set to obtain a hierarchical group model;

[0008] Step 40, generate multiple candidate models according to the component feature set with the largest posterior probability, based on the hierarchical group model, by using the Bayesian rule again, obtain the confidence intervals of all candidate models, combine the accuracy of each candidate model, and screen out the octane number and cetane number models suitable for the actual situation of the refinery according to the principle of multi-objective optimization.

[0009] Furthermore, in step 20, the component feature set with the largest posterior probability is screened, specifically including:

[0010] A single model m belongs to a set of candidate models M. Each model follows the distribution of a known data set Y. f(y|m,β m ), where the parameter vector β m ∈B m , B m is the set of possible values ​​of the coefficients of model m. Assuming the prior probability of model m is f(m), the posterior probability is:

[0011]

[0012] Among them, f(y|m) is the edge similarity, which is given by f(y|m)=∫f(y|m,β m )f(β m |m)dβ m and f(β m |m) is calculated and its value is estimated approximately by using the Markov Monte Carlo random sampling method. The sampling range is (m,β m ) Space:

[0013]

[0014] in, Where p is the total number of features.

[0015] Furthermore, the step 30 includes a data preprocessing process and a modeling verification process;

[0016] The data preprocessing process is: normalizing the data set by using a probability statistics method, and then directly performing cluster analysis on the data set by using an unsupervised learning method, approximating the sparse holes in the feature space of the data set, and obtaining a training set;

[0017] In the modeling verification process, the modeling of hierarchical groups adopts a linear cumulative function, and the formula is as follows:

[0018]

[0019] Among them, function f(Y) is the function of the property to be predicted, C i is the contribution of the i-th group in the primary group, N i is the number of occurrences of group i, δ is the primary group coefficient; w is the secondary group coefficient, D j is the contribution of the j group in the secondary group, M j is the number of occurrences; λ is the group coefficient of the third-level component descriptor, and f(Y*) is the total contribution of the third-level descriptor to a given property;

[0020] Calculate the hierarchical group coefficients δ, w, λ and group contribution C i , D j When the hierarchical method is used to regress in sequence, C is obtained by regressing the training set. i ; Then regression is performed to obtain the secondary group contribution D j ; f(Y*) is calculated by the component descriptor without regression calculation, and finally unified regression is used to obtain the group coefficients δ, w, λ, that is, the size of the weight, which represents the influence of the group fragment at the corresponding level on the given property.

[0021] The present invention has the following advantages:

[0022] By selecting radical fragments based on the mechanism and combining them with component descriptors that are independent of data regression coefficients, the number of required regression calculation coefficients is reduced, which reduces the dependence on the size of the data set to a considerable extent. At the same time, the posterior distribution probability of the feature subset model is given to achieve "soft" constraints, which is suitable for predictive research on the refining properties of pure components with limited data. On this basis, the Bayesian rule is introduced again to estimate the posterior probability distribution of the final model and avoid the risk of overfitting of the final prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The present invention will be further described below in conjunction with embodiments with reference to the accompanying drawings.

[0024] Figure 1 It is a schematic diagram of the process of the present invention;

[0025] Figure 2 Schematic diagram of hierarchical groups of the present invention;

[0026] Figure 3 A schematic diagram of a screening process for constructing a component feature set of the present invention;

[0027] Figure 4 Schematic diagram of the modeling process of the hierarchical group of the present invention;

[0028] Figure 5 This is one of the schematic diagrams of the uncertainty analysis process of the candidate model of the present invention;

[0029] Figure 6 This is the second schematic diagram of the uncertainty analysis process of candidate models of this method. DETAILED DESCRIPTION

[0030] The embodiment of the present invention provides a method for predicting the refining properties of pure components based on hierarchical group construction, and predicts the octane number of pure components in petroleum products, such as the octane number of each component in gasoline products and the cetane number of each component in diesel products. The component feature set is hierarchically constructed, and when the component descriptors are added to form a feature set, the Bayesian rule is introduced to estimate the posterior probability distribution thereof. On this basis, the hierarchical group construction is introduced, and the group fragments are hierarchically constructed to avoid the risk of overfitting in the final prediction.

[0031] like Figure 1 As shown, the overall idea of ​​the present invention is as follows:

[0032] S1: Construct a predefined group fragment component library.

[0033] In view of the defects of the existing group fragment construction method, a new set of component feature sets is proposed to characterize the refining properties of components in petroleum products. This component feature set combines the characteristic groups of the mechanism with the component descriptors screened by machine learning to characterize the refining properties of the components. While constructing the groups and descriptors, they are divided into levels, and the higher the level of the group, the more detailed the description of the component.

[0034] The primary group contains the basic groups of the component structure, such as -CH, -CH3, -CO, etc. Simple structure components such as alkanes can be disassembled and characterized through this hierarchical group. However, this hierarchical group can only represent the basic composition of the component, and cannot characterize the linking position of the group in the component. The different linking positions have a significant impact on the refining properties of the component.

[0035] Therefore, the secondary groups focus on building blocks, that is, combinations of basic groups, to distinguish aromatics from alkanes and their corresponding isomers. Figure 2 As shown, the primary basic group includes the R group representing the aromatic ring A6 group and CH2. The -CH2 group is linked to the benzene ring, so it forms a new group block aC-CH with the carbon on the benzene ring to which it is connected, which is represented in the secondary group block to characterize the component.

[0036] The tertiary groups use component descriptors. Due to the large number of component descriptors and the accuracy of descriptors based on quantum chemical calculations, there is still some controversy in the scientific community. Therefore, we will focus on descriptors that describe the topological structure of components, such as the connectivity index.

[0037] S2. Construct and screen component feature sets based on target components.

[0038] like Figure 3 As shown, when the refining properties of the target component need to be predicted, the codable, simplified component expression method SMILES is used to represent the complex component structure with a two-dimensional code, and the molecular structure of the given component is automatically decomposed into group fragments that conform to the component library, so as to perform quantitative analysis.

[0039] First, primary and secondary groups are screened from the group library according to the molecular structure of the target component. Global optimization algorithms such as simulated annealing and genetic algorithms can be used for screening. Then, tertiary groups are added to the screened feature set. However, the addition of tertiary groups will inevitably overlap with primary and secondary groups, resulting in feature set redundancy. Therefore, combining information theory with machine learning, the concept of minimum correlation-maximum information is introduced to ensure that the added tertiary groups maintain the minimum correlation with the existing low-level groups, while maintaining the maximum information with the properties to be predicted, that is, to maximize the characterization of the properties to be predicted.

[0040] Then, the Bayesian rule is introduced to perform feature selection and calculate the posterior probability of the candidate model. A single model m belongs to the candidate model set M. Each model obeys the distribution of the known data set Y. f(y|m,β m ), where the parameter vector β m ∈B m , B m is the set of possible values ​​of the coefficients of model m. Let the prior probability of model m be f(m), then the posterior probability is:

[0041]

[0042] Among them, f(y|m) is the edge similarity, which can be expressed as f(y|m)=∫f(y|m,β m )f(β m |m)dβ m and f(β m |m). However, this integral cannot be solved analytically in most cases, so the Markov Monte Carlo (MCMC) random sampling method is used to approximate its value. The sampling range is (m, β m ) Space:

[0043]

[0044] Feature selection is a branch problem for model selection, that is, using binomial distribution to represent candidate models. Where p is the total number of features. Thus, the posterior distribution probability of the model represented by each feature subset based on the known data set Y is obtained, thereby achieving "soft" constraints. The core of this Bayesian rule-based feature selection method is the MCMC sampling method.

[0045] By selecting group fragments based on the mechanism and combining them with component descriptors that are independent of data regression coefficients, the number of required regression calculation coefficients is reduced, which reduces the dependence on the size of the data set to a considerable extent. At the same time, the posterior distribution probability of the feature subset model is given to achieve "soft" constraints, which is suitable for predictive research on the refining properties of pure components with limited data.

[0046] S3: Perform hierarchical group modeling and coefficient solution.

[0047] The hierarchical group modeling and coefficient solution process is as follows Figure 4 As shown in the figure, the overall process can be divided into two parts: data preprocessing and modeling verification. Due to the sparseness of the existing data set of component refining properties, advanced statistical and machine learning methods need to be introduced in the data preprocessing stage to improve the accuracy of small sample data modeling.

[0048] The distribution of the characteristic values ​​and experimental values ​​of the components in the database is difficult to meet the requirements of normal distribution, which will affect the model effect during the modeling process. It is necessary to use the probability statistics method, namely the Box-Cox log-likelihood function method, to transform it into a normal state. Due to the sparsity of the feature space, the randomly selected training set is difficult to cover the feature space of the test set, resulting in excessive extrapolation of the model based on the training set, which reduces the model prediction effect. Therefore, the second step adopts an unsupervised learning method, that is, only the feature set of the data set is targeted without evaluating the modeling effect, and the data set is directly clustered. The sparse holes in the feature space in the data set are approximately estimated. Based on this, the selected training set can cover the feature space of the test set samples to the greatest extent, thereby improving the model prediction effect.

[0049] The modeling of hierarchical groups gives priority to the traditional linear cumulative function because it has a small amount of calculation and can give the contribution coefficient of the corresponding group, which provides richer mechanism information to a certain extent. Its formula is shown as follows:

[0050]

[0051] Among them, function f(Y) is the function of the property to be predicted, C i is the contribution of the i-th group in the primary group, N i is the number of occurrences of group i, δ is the primary group coefficient; w is the secondary group coefficient, D jis the contribution of the j group in the secondary group, M j is the number of occurrences; λ is the group coefficient of the third-level component descriptor, and f(Y*) is the total contribution of the third-level descriptor to a given property.

[0052] Calculate the hierarchical group coefficients δ, w, λ and group contribution C i , D j When , a hierarchical method is used to regress in sequence. C is obtained by regression of the training set i ; Then regression is performed to obtain the secondary group contribution D j ; Since f(Y*) is calculated from the component descriptor, no regression calculation is required, which greatly reduces the demand for the size of the training set. Finally, the group coefficients δ, w, λ are obtained by unified regression. The calculated group coefficients δ, w, λ, that is, the size of the weight, can represent the influence of the group fragment at the corresponding level on the given property.

[0053] S4: Perform uncertainty analysis.

[0054] like Figure 5 and Figure 6 As shown in the figure, the uncertainty analysis of the predicted value, i.e. the estimation of the confidence interval, is crucial to the practical application of the model. Since the hierarchical group model has an explicit mathematical expression and includes the probability distribution of each candidate model, the confidence interval of all candidate models can be obtained by using the Bayesian rule again. Combined with the accuracy of each model, the accuracy and practicality of the model are comprehensively considered to obtain an octane number and cetane number model that is more suitable for the actual situation of the refinery.

[0055] It should be noted that, in the process of performing yield rate prediction calculation, relevant technical personnel in the field can make appropriate deformations and corresponding parameter settings according to relevant principles. The above-mentioned embodiment only expresses one implementation of the present invention, and its description is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention patent.

[0056] A specific embodiment of the present invention is as follows:

[0057] Step 10, using the codable simplified component expression SMILES, the complex component structure is represented by two-dimensional coding, and a predefined group fragment component library is constructed, including primary groups, secondary groups and tertiary groups; the primary groups are basic groups containing component structures; the secondary groups are the link position combinations of the basic groups, which are used to distinguish aromatics from alkanes and the corresponding isomers; the tertiary groups are descriptors that describe the component topological structure;

[0058] Step 20, according to the molecular structure of the target component, a primary group and a secondary group are screened out from the predefined group fragment component library, and then according to the refining property to be predicted of the target component, a plurality of tertiary groups are screened out by using the principle of maintaining the minimum correlation with the primary group and the secondary group and maintaining the maximum amount of information with the property to be predicted, and any number of tertiary groups are randomly selected to form a plurality of component feature sets with the primary group and the secondary group, and then the component feature set with the largest posterior probability is screened out;

[0059] Step 30, using a linear cumulative function to combine groups at different levels for modeling and then solving the coefficients through the training set to obtain a hierarchical group model;

[0060] Step 40, generate multiple candidate models according to the component feature set with the largest posterior probability, based on the hierarchical group model, by using the Bayesian rule again, obtain the confidence intervals of all candidate models, combine the accuracy of each candidate model, and screen out the octane number and cetane number models suitable for the actual situation of the refinery according to the principle of multi-objective optimization.

[0061] In step 20, the component feature set with the largest posterior probability is screened, which specifically includes:

[0062] A single model m belongs to a set of candidate models M. Each model follows the distribution of a known data set Y. f(y|m,β m ), where the parameter vector β m ∈B m , B m is the set of possible values ​​of the coefficients of model m. Assuming the prior probability of model m is f(m), the posterior probability is:

[0063]

[0064] Among them, f(y|m) is the edge similarity, which is given by f(y|m)=∫f(y|m,β m )f(β m |m)dβ m and f(β m |m) is calculated and its value is estimated approximately by using the Markov Monte Carlo random sampling method. The sampling range is (m,β m ) Space:

[0065]

[0066] in, Where p is the total number of features.

[0067] The step 30 includes a data preprocessing process and a modeling verification process;

[0068] The data preprocessing process is: normalizing the data set by using a probability statistics method, and then directly performing cluster analysis on the data set by using an unsupervised learning method, approximating the sparse holes in the feature space of the data set, and obtaining a training set;

[0069] In the modeling verification process, the modeling of hierarchical groups adopts a linear cumulative function, and the formula is as follows:

[0070]

[0071] Among them, function f(Y) is the function of the property to be predicted, C i is the contribution of the i-th group in the primary group, N i is the number of occurrences of group i, δ is the primary group coefficient; w is the secondary group coefficient, D j is the contribution of the j group in the secondary group, M j is the number of occurrences; λ is the group coefficient of the third-level component descriptor, and f(Y*) is the total contribution of the third-level descriptor to a given property;

[0072] Calculate the hierarchical group coefficients δ, w, λ and group contribution C i , D j When the hierarchical method is used to regress in sequence, C is obtained by regressing the training set. i ; Then regression is performed to obtain the secondary group contribution D j ; f(Y*) is calculated by the component descriptor without regression calculation, and finally unified regression is used to obtain the group coefficients δ, w, λ, that is, the size of the weight, which represents the influence of the group fragment at the corresponding level on the given property.

[0073] The present invention selects radical fragments based on the mechanism and combines them with component descriptors that are independent of data regression coefficients, thereby reducing the number of required regression calculation coefficients and, to a considerable extent, reducing the dependence on the size of the data set. At the same time, the posterior distribution probability of the feature subset model is given to achieve "soft" constraints, which is suitable for predictive research on the refining properties of pure components with limited data. On this basis, the Bayesian rule is introduced again, so that the posterior probability distribution of the final model can be estimated to avoid the risk of overfitting of the final prediction model.

[0074] Although the specific implementation modes of the present invention are described above, those skilled in the art should understand that the specific implementation modes described are only illustrative and are not intended to limit the scope of the present invention. Equivalent modifications and changes made by those skilled in the art in accordance with the spirit of the present invention should be included in the scope of protection of the claims of the present invention.

Claims

1. A method for predicting the refining properties of pure components based on hierarchical group construction, characterized in that: include: Step 10, using the codable simplified component expression SMILES, the complex component structure is represented by two-dimensional coding, and a predefined group fragment component library is constructed, including primary groups, secondary groups and tertiary groups; the primary groups are basic groups containing component structures; the secondary groups are the link position combinations of the basic groups, which are used to distinguish aromatics from alkanes and the corresponding isomers; the tertiary groups are descriptors that describe the component topological structure; Step 20, according to the molecular structure of the target component, a primary group and a secondary group are screened from the predefined group fragment component library, and then according to the refining property to be predicted of the target component, a plurality of tertiary groups are screened by using the principle of maintaining the minimum correlation with the primary group and the secondary group and maintaining the maximum amount of information with the property to be predicted, and any number of tertiary groups are randomly selected to form a plurality of component feature sets with the screened primary group and the secondary group, and then the component feature set with the largest posterior probability is screened; Step 30, using a linear cumulative function to combine groups at different levels for modeling and then solving the coefficients through the training set to obtain a hierarchical group model; Step 40, generate multiple candidate models according to the component feature set with the largest posterior probability, based on the hierarchical group model, by using the Bayesian rule again, obtain the confidence intervals of all candidate models, combine the accuracy of each candidate model, and screen out the octane number and cetane number models suitable for the actual situation of the refinery according to the principle of multi-objective optimization.

2. The method according to claim 1, characterized in that: In step 20, the component feature set with the largest posterior probability is screened, which specifically includes: Single Model m Belongs to the candidate model set M Each model is subject to a known data set Y The distribution of f ( y | m , β m ), where the parameter vector β m ∈ B m , B m For Model m The set of possible values ​​of the coefficients, let the model m The prior probability of f ( m ), then the posterior probability is: in, f ( y | m ) is the edge similarity, and Calculated, the value is approximately estimated using the Markov Monte Carlo random sampling method, and the sampling range is ( m , β m )Space: in, , where p is the total number of features.

3. The method according to claim 1, characterized in that: The step 30 includes a data preprocessing process and a modeling verification process; The data preprocessing process is: normalizing the data set by using a probability statistics method, and then directly performing cluster analysis on the data set by using an unsupervised learning method, approximating the sparse holes in the feature space of the data set, and obtaining a training set; In the modeling verification process, the modeling of hierarchical groups adopts a linear cumulative function, and the formula is as follows: Among them, the function f ( Y ) is a function of the property to be predicted, C i The first i The contribution of the group, N i for i The number of group occurrences, δ is the primary group coefficient; w is the secondary group coefficient, D j For the secondary group j The contribution of the group, M j is the number of times it occurs; λ is the third-level component descriptor group coefficient, f ( Y* ) is the total contribution of the three-level descriptors to a given property; Calculate the hierarchical group coefficient δ , w , λ and group contribution C i , D j When , the hierarchical method is used to regress in sequence, and the training set regression is obtained C i ; Then regression is performed to obtain the secondary group contribution D j ; f ( Y* ) is calculated from the component descriptor, without regression calculation, and finally the group coefficient is obtained by unified regression δ , w , λ , that is, the size of the weight, represents the influence of the group fragment at the corresponding level on the given property.

Images